Pull all PR review comments (any reviewer, human or bot), classify each as convergent (a fix) or divergent (a design disagreement), apply convergent ones after sign-off, escalate divergent ones to the author as questions. The single path for responding to a PR review; wires /continue step 11.
Enforces the required_agents declared by matched policies at runtime. Run at step 5 of /continue and as a self-check before any Edit/Write, to verify the right expert agents are in context for the files touched. Matches policies by triggered_by globs; blocks until every missing agent is loaded.
For an ad-hoc dev task with no PRD — brainstorm the approach, converge on a short written spec, and get explicit approval BEFORE writing code. A lightweight design-before-code gate. Triggers on "let's design this", "how should I approach", "brainstorm this".
Break a feature / ClickUp parent task / design doc into one or more GitHub issues with the BeWith 6-section template + AI-assigned effort/urgency/domain labels. Triggers on phrases like "break down this task", "create issues for this design", "פצל את המשימה הזאת ל-issues".
Bring the local BeWith stack up and confirm it is healthy — sandbox infra (Mongo/MySQL/Redis/LocalStack) plus the NestJS services (headless "Debug-Watch All Services") — then optionally exercise a change end-to-end. Idempotent; the hands-on counterpart to /verify.
Initialize an existing bewith-dev repo onto the platform — scaffold CLAUDE.md + the CI caller, retire legacy .cursor rules (archive, then rewrite repo-specific ones into CLAUDE.md), verify the plugin + git hooks load. Worktree-isolated; prepares a reviewable branch, never merges.
Opens a CTO-review ticket in a dedicated ClickUp list, carrying the full handoff package (problem, UI composition, design reference, acceptance criteria, links). Title convention [Feature name] informative title, tag cto review. Runs at the end of the product flow or manually off-flow.
Capture a discussion that reached decisions into a durable record — an ADR (technical) or a decision log entry (product/business) — then offer to break its action items into tasks via /breakdown. Worktree-isolated. Triggers on "record this decision", "תעד את ההחלטה".
Decide when and how to push token-heavy operations into a subagent (or use cheaper tool alternatives) so only a short verdict returns to the main thread. Implements the subagent-as-context-firewall lever from context-and-cost-economy.md.
Turn a screen/feature requirement into a concrete UI composition from the design system — which COMPONENTS.md components, layout, states, and tokens — the machine-consumable design reference when there's no Figma (Path B). Flags gaps as /add-component; never invents components or raw hex.
Walks Claude through a session wrap-up — lists PRs touched, scans memory for stale references, and writes or confirms memory entries before the session closes. Invoked manually via /end-session, or by Claude when the developer signals closure ("done", "thanks", "סיימנו").
Favorite/pin a Claude Code session so you can find and reopen it later. Pins the current session (by $CLAUDE_CODE_SESSION_ID) to a private GitHub gist so favorites follow you across machines; lists, opens, and removes them. Invoked by /favorite, /favorites, /open-favorite, /unfavorite.
Identify and document a bug from inputs (description, logs, screenshots, JAM links, a ClickUp task) for someone else to fix — without implementing the fix unless asked. Gathers inputs, traces the flow, finds root cause, writes a structured understanding doc. Called by qa-engineer.
Render a finished incident / investigation into a browser-viewable RTL-Hebrew HTML report artifact (via the Artifact tool) — root cause, evidence, impact, timeline, fix. Use after investigating, not for finding bugs (see identify-bug).
Capture a what's-missing / could-change POC observation into docs/insights/ — fast, separate from proposals. Worktree-isolated. Invoked by /insight or when the developer voices a gap or possible change during the build. Does NOT create a proposal; the 2-3 day review promotes agreed insights.
Feature-level release-readiness before shipping a whole feature to production — flag + kill-switch, staged rollout, rollback plan, release notes/version, monitoring + post-deploy watch, comms, success metrics. Broader than the per-PR DoD. Produces a checklist; gates nothing automatically.
Create version-controllable software diagrams with Mermaid's text syntax — flowcharts, sequence, class, ERD, C4/architecture, state. Use when documenting architecture, a data model, an API/auth flow, or a DB schema, or any "diagram / visualize / map out the flow" request.
Rank my open ClickUp tasks + assigned GitHub Issues by priority → urgency → effort and recommend what to start next; --fits-in Nm/Nh/Nd filters to what fits a time box. Read-only; the ranking is computed in a script, not by the model. Invoked by /next-task.
Save an added or changed HTTP endpoint in the organization R&D Postman workspace with a sample request and a saved response example, at the end of development. The Definition-of-Done step for API work — keeps the collection an accurate, runnable contract. Never commits secrets.
Recover lost or clobbered code from the automatic safety snapshots — falls back to git reflog/stash/dangling blobs. Triggered by /restore, free language ("I lost code"), or proactively when Claude detects dropped code. Diffs and confirms before writing; reversible.
Read the local per-machine session registry and render a cross-session mission-control view — branch, cwd, files, age, live/stale, collisions, and which sessions are waiting on the developer. Optionally launches a self-refreshing HTML dashboard. Read-only; invoked by /sessions.
Keeps a task's state mirrored across both trackers — writes the AI Context block to the GitHub Issue and a condensed mirror to the ClickUp ai_context field, and maintains the bidirectional link so any session knows which ClickUp task an Issue belongs to. Runs at checkpoints; never gated.
A 4-phase active-fix loop — reproduce, isolate, hypothesize+instrument, fix+verify — with a hard stop after 3 failed attempts. Instruments at the gateway/RPC/service/Mongo boundaries. Triggers on "debug this", "why is this failing", "help me fix this bug".
Drive a change test-first — a failing test, then minimum code to pass, then refactor. Adapted to the bewith Jest/NestJS stack. The Iron Law — no production code before a failing test. Triggers on "TDD this", "test first", "drive this test-first".
After a merge, watch the deployment to its target environment and the observability stack for a defined window before the task closes. Confirms the deploy succeeded and telemetry is clean (vendor-neutral); on a regression it flags an incident instead of closing the task.